Optimization of neural network configurations for short-term traffic flow forecasting using orthogonal design

Neural networks have been applied for short-term traffic flow forecasting with reasonable accuracy. Past traffic flow data, which has been captured by on-road sensors, is used as the inputs of neural networks. The size of this data significantly affects the performance of short-term traffic flow forecasting, as too many inputs result in over-specification of neural networks and too few inputs result in under-learning of neural networks. However, the amount of past traffic flow data input, is usually determined by the trial and error method. In this paper, an experimental design method, namely orthogonal design, is used to determine appropriate amount of past traffic flow data for neural networks for short-term traffic flow forecasting. The effectiveness of the orthogonal design is demonstrated by developing neural networks for short-term traffic flow forecasting based on past traffic flow data captured by on-road sensors located on a freeway in Western Australia.